Physical Attribute Atlas-Driven Machine Learning Model for Co-seismic Landslide Prediction

Earthquake-induced landslide susceptibility prediction (ELSP) provides a critical scientific basis for post-earthquake emergency response and reconstruction planning. A fundamental challenge lies in the transient characterization of slope dynamic stability under strong ground motion. To address the strong dependence of conventional approaches on observed landslide inventories and in situ monitoring data, this study develops a Physical Attribute Atlas-driven Machine Learning (PAA-ML) model for co-seismic landslide prediction. First, a Python-driven automated numerical simulation platform is established to perform large-scale parallel analyses of GeoStudio models, enabling the construction of a high-dimensional Physical Attribute Atlas that integrates ground motion intensity measures (IMs), geomechanical properties of geomaterials, and slope geometric characteristics. This atlas systematically characterizes the coupled relationships among multi-source physical factors and their governing mechanisms on dynamic slope responses. Subsequently, the Light Gradient Boosting Machine (LightGBM) algorithm is employed to capture complex nonlinear mappings within the high-dimensional feature space, enabling accurate prediction of permanent slope displacement and instability probability. Empirical results show that incorporating multidimensional physical attributes enhances the mechanistic correlation of the data-driven framework, ensuring robust performance and high generalizability despite initial post-earthquake lags in high-dimensional ground motion data. Relying solely on basic topography and seismic inputs, the PAA-ML model enables quantitative estimation of permanent displacements and rapid probabilistic hazard mapping. This study provides an efficient, reliable computational paradigm for co-seismic landslide risk assessment in tectonically complex regions, offering valuable insights into integrating physics-priors with data-driven methods.